Segmentation of Tuberculosis Bacilli in Conventional Microscopy Images Through Accelerated CNN Using Dilated Convolutions

Philipe Rangel Demuth, Patrick Marques Ciarelli, Jorge Leonid Aching Samatelo


In this work, we propose a method to achieve microscopy image segmentation,
in which a convolutional neural network (CNN) is used. The method is divided in two parts: (i) the CNN is trained for pixelwise classification of image; (ii) the training CNN is accelerated, removing the redundant operations, allowing the classification of the pixels from an entire image patch at the same time. The method was evaluated over a dataset with 120 images obtained using conventional microscopy in sputum smear sheets prepared according to the Ziehl-Neelsen technique. In the experimental evaluations carried out on this dataset, we obtained an accuracy of 97:33% and recall of 96:30%. The accelerated CNN is 44 times faster, maintaining identical prediction results. These results show that the proposed method has the potential to handle the given problem.

Texto completo:




  • Não há apontamentos.